app.py

import pandas as pd
import faiss
from sentence_transformers import SentenceTransformer
from transformers import pipeline
import streamlit as st

# Load models
@st.cache_resource
def load_embedding_model():
    return SentenceTransformer('all-MiniLM-L6-v2')  # Using Sentence-BERT

@st.cache_resource
def load_text_generator():
    return pipeline("text2text-generation", model="t5-small")

@st.cache_resource
def load_rephrasing_model():
    return pipeline("text2text-generation", model="t5-base")

embedding_model = load_embedding_model()
text_generator = load_text_generator()
rephrasing_model = load_rephrasing_model()

# Preprocess and embed data
def preprocess_and_embed(data):
    text_data = data.astype(str).apply(" ".join, axis=1)  # Concatenate rows
    embeddings = embedding_model.encode(text_data.tolist(), convert_to_numpy=True)
    return text_data, embeddings

# Create FAISS index
def create_faiss_index(embeddings):
    dimension = embeddings.shape[1]
    index = faiss.IndexFlatL2(dimension)
    index.add(embeddings)
    return index

# Retrieve recommendations
def retrieve_recommendations(query, text_data, index, embeddings, k=5):
    query_embedding = embedding_model.encode([query], convert_to_numpy=True)
    distances, indices = index.search(query_embedding, k)
    results = [{"text": text_data[i], "distance": d} for i, d in zip(indices[0], distances[0])]
    return results

# Generate contextual response based on query and context
def generate_response(query, recommendations):
    context = " ".join([rec["text"] for rec in recommendations])
    limited_context = f"Context: {context}\nQuery: {query}\nAnswer:"
    response = text_generator(limited_context, max_length=100, num_beams=3)
    return response[0]['generated_text']

# Rephrase the generated response
def rephrase_response(response):
    rephrased = rephrasing_model(f"Rephrase this: {response}", max_length=100, num_beams=3)
    return rephrased[0]['generated_text']

# Streamlit UI
st.title("RAG-Based AI Recommendation System (Using SBERT)")

uploaded_file = st.file_uploader("Upload CSV/Excel file", type=["csv", "xlsx"])

if uploaded_file:
    file_type = uploaded_file.name.split(".")[-1]
    if file_type == "csv":
        data = pd.read_csv(uploaded_file)
    else:
        data = pd.read_excel(uploaded_file)

    st.write("Data Preview:")
    st.write(data.head())

    # Preprocess and embed
    with st.spinner("Processing data and creating embeddings..."):
        text_data, embeddings = preprocess_and_embed(data)
        index = create_faiss_index(embeddings)
        st.success("Data processed and embeddings created!")

    # Query input
    query = st.text_input("Enter your query:")

    if query:
        # Retrieve recommendations
        with st.spinner("Retrieving recommendations..."):
            recommendations = retrieve_recommendations(query, text_data, index, embeddings)
            st.write("Top Recommendations:")
            for rec in recommendations:
                st.write(f"- {rec['text']} (Distance: {rec['distance']:.2f})")

        # Generate contextual response
        with st.spinner("Generating contextual recommendation..."):
            refined_response = generate_response(query, recommendations)

        # Rephrase the response
        with st.spinner("Rephrasing the recommendation for better grammar and clarity..."):
            final_response = rephrase_response(refined_response)
            st.write("Contextual Recommendation:")
            st.write(final_response)